Estimation and forecasting for binomial and negative binomial INAR(1)

Estimation and forecasting for binomial and
negative binomial INAR(1) time series
Stima e previsione di modelli INAR(1) con innovazioni
Binomiali e Binomiali Negative
Luisa Bisaglia, Margherita Gerolimetto and Paolo Gorgi
Abstract To model count time series Al-Osh & Alzaid (1987) and McKenzie (1988)
introduced the INteger-valued AutoRegressive process (INAR). Usually the innovation term is assumed to follow a Poisson distribution. However, other distributional
assumptions may be used instead. In this work we discuss the issue of estimating and
forecasting in case of INAR(1) time series with over and under-dispersion, resorting respectively to the Binomial and Negative Binomial distributions. We calculate
the maximum likelihood functions for the considered cases and via a Monte Carlo
experiment we show that the resulting estimators have a good performance. Moreover, we also concentrate on the problem of producing coherent predictions based
on estimates of the p-step ahead predictive mass functions assuming Binomial and
Negative Binomial distributions of the error term. Finally, we compare the forecast
accuracy of Binomial and Negative Binomial INAR with that of Poisson INAR and
ARMA models with a Monte Carlo experiment.
Abstract Per modellare serie storiche per dati di conteggio Al-Osh & Alzaid (1987)
and McKenzie (1988) hanno introdotto i processi AR a valori interi (INAR). In generale, la distribuzione del termine d’errore si assume essere Poisson, tuttavia, allo
scopo di poter trattare adeguatamente anche casi di sotto e sovra-dispersione, in
questo lavoro verranno considerate le distribuzioni Binomiale e Binomiale Negativa. Pi`u specificamente, sotto queste ipotesi distribuzionali, verranno dapprima
sviluppate le procedure di stima e previsione, quindi, attraverso un esperimento di
Monte Carlo, ne verranno valutate le relative performance.
Key words: COUNT TIME SERIES, INAR(1) PROCESS, FORECASTING
L. Bisaglia
Dept. of Statistical Science, University of Padua, e-mail: [email protected]
M. Gerolimetto
Dept. of Economics, University of Ca’ Foscari, Venice, e-mail: [email protected]
P. Gorgi
Dept. of Statistical Science, University of Padua, e-mail: [email protected]
1
2
Luisa Bisaglia, Margherita Gerolimetto and Paolo Gorgi
1 Introduction
Recently, there has been a growing interest in studying nonnegative integer-valued
time series and, in particular, time series of counts. The most common approach to
build an integer-valued autoregressive processes is based on a probabilistic operation called (binomial) thinning, as reported in Al-Osh & Alzaid (1987) and McKenzie (1988) who first introduced integer-valued autoregressive processes (INAR).
While theoretical properties of INAR models have been extensively studied in the
literature, relatively few contributions discuss the development of forecasting methods that are coherent, in the sense of producing only integer forecasts of the count
variable. Freeland & McCabe (2004), in the context of INAR(1) process with Poisson innovations suggest the use of the median of the k−step-ahead conditional distribution to emphasize the intention of preserving the integer structure of the data in
generating the forecasts. The Poisson distribution, however, has the disadvantage to
take into account only equi-dispersion. Thus, unlike the usual applications where the
error terms are usually Poisson, in oder to investigate also over and under-dispersion
cases, we assume the error terms to follow the Binomial or Negative Binomial distributions.
The purpose of this work is, firstly, to calculate the maximum likelihood (ML)
functions and consequently the estimators of the parameters of the Binomial and
Negative Binomial INAR(1) (respectively BINAR(1) and BNINAR(1)) models, secondly, to obtain the predictive probability mass function (pmf) in order to derive
coherent predictions for the considered models.
The remainder of the paper is divided as follows. Section 2 recalls INAR(1)
model. Section 3 provides the theoretical results in order to obtain the ML function and the predictive pmf. Section 4 describes our simulation study to show the
performance of the estimators and to compare the forecast accuracy of BINAR and
BNINAR models with that of Poisson INAR (PoINAR) and ARMA ones.
2 The INAR(1) model
To introduce the class of INAR model we first recall the thinning operator, ‘◦’,
defined as follows.
Definition Let Y be a non negative integer-valued random variable, then for any
α ∈ [0, 1]
Y
α ◦Y = ∑ Xi
i=1
where Xi is a sequence of iid count random variables, independent of Y, with common mean α.
The INAR(1) process {Yt ;t ∈ Z} is defined by the recursion
Estimation and forecasting for binomial and negative binomial INAR(1) time series
Yt = α ◦Yt−1 + εt
3
(1)
where α ∈ [0, 1], and εt is a sequence of iid discrete random variables with finite first
and second moment. The components of the process {Yt } are the surviving elements
of the process Yt−1 during the period (t − 1,t], and the number of elements which
entered the system in the same interval, εt . Each element of Yt−1 survives with probability α and its survival has no effect on the survival of the other elements, nor on
εt which is not observed and cannot be derived from the Y process in the INAR(1)
model. If the error terms are distributed as a Po(λ ), the marginal distribution of the
observed counts is a Poisson distribution with (unconditional) mean and variance
equal to λ /(1 − α) (Jung et al., 2005). Clearly, this model does not allow for under
or over-dispersion in data and for this reason in the next Section we will consider
Binomial and Negative Binomial distribution instead.
3 Theoretical results
Theoretical properties of PoINAR models have been discussed extensively in the
literature. In this Section we extend the results of the PoINAR(1) model to the BINAR(1) and NBINAR(1) models.
3.1 Maximum Likelihood function for BINAR(1) and NBINAR(1)
We begin assuming that the distribution of the innovation term is Bi(m, p). Differently from the PoINAR model, now the marginal distribution of Yt is not known and
the distribution of Y0 could not be obtained explicitly. Thus we calculate the conn
p(Yt |Yt−1 ) and the log-likelihood
ditional likelihood function L(α, m, p|y0 ) = ∏t=1
that is:
n
l(α, m, p|y0 ) = log(L(α, m, p|y0 )) =
∑ log(p(Yt |Yt−1 )) =
t=1
n
!
n
!
p
+
= ∑ yt−1 log(1 − α) + mn log(1 − p) + ∑ yt log
(1 − p)
t=1
t=1
!
Min(yt ,yt−1 ) n
yt−1
m
α(1 − p) i
+ ∑ log
∑
i
yt − i
p(1 − α)
t=1
i=Max(0,y −m)
(2)
t
The ML estimate are the values of α, m and p that maximizes the (2).
Now, we assume that the distribution of the innovation term is NB(r, p). In this
case it is possible to show that the conditional log-likelihood is:
4
Luisa Bisaglia, Margherita Gerolimetto and Paolo Gorgi
n
l(α, r, p|y0 ) = log(L(α, r, p|y0 )) =
∑ log(p(Yt |Yt−1 )) =
t=1
!
n
=
∑ yt−1
log(1 − α) + nr log(p) +

min(yt ,yt−1 ) + ∑ log 
t=1
∑ yt
log(1 − p) +
t=1
t=1
n
!
n
∑
i=0
yt−1 Γ (r + yt − i)
i
Γ (r)(yt − i)!
α
(1 − p)(1 − α)
!i 

(3)
The ML estimates are the values of α, r and p that maximizes the (3).
3.2 k-step ahead predictive probability mass function
Freeland & McCabe (2004) define the p−step ahead predictive probability mass
function (pmf) for the PoINAR(1) model as:
min(y,YT ) YT k s
k YT −s ×
s (α ) (1 − α )
n
o
y−s
1−α k
1−α k
1
exp
−λ
×
λ
1−α
1−α
(y−s)!
Pk (YT +k = y | YT ) = ∑s=0
(4)
Consequently, Pk (YT +k = y | YT ) is the probability of the value y occurring, according
to the k−step-ahead conditional distribution. In order to obtain coherent predictions
for YT +k , Freeland & McCabe (2004) suggest using the median of the k−step-ahead
ˆ λˆ ), so typically (α, λ ) are estipmf. In practice we compute Pk (YT +k = y | YT , α,
mated via ML.
Unlike the PoINAR model, for the BINAR and BNINAR models it is not possible
to derive a general form for the p−step ahead predictive probability mass function.
However, it is possible to calculate it for some values of p and in particular, we obtain the expression of the conditional probabilities for BINAR and BNINAR models
in case of p=1, 2. Here, for space reasons, we report only the expression relatively
to the case of p=1 for the BNINAR:
yt r
yt+1
p(Yt+1 |Yt ) = (1 − α) p (1 − p)
min(yt+1 ,yt ) ∑
i=0
Γ (r + yt+1 − i)
×
Γ (r)(yt+1 − i)!
!
yt
i
α
(1 − p)(1 − α)
!i
(5)
Estimation and forecasting for binomial and negative binomial INAR(1) time series
5
4 Monte Carlo experiments
In this Section we provide the details of the Monte Carlo experiment we carried
out. We generate data from INAR(1) DGP’s with Poisson errors, Binomial errors and Negative Binomial errors for different values of parameters and sample
sizes. For each model we generated s = 1000 independent realizations. To compare
the forecasting performance of the estimated models, we computed the Forecast
Mean Square Error (FMSE) and Forecast Mean Absolute Error (FMAE) statistics
of k−step-ahead forecasts, where k = 1, 2, and the Total variation and Bhattacharya
distance (BC) between the true distribution h− steps ahead and the predicted one.
For lack of space we report only some results for the BNINAR(1) case. As we
can notice from Table 1 and Figure 1, the Negative Binomial hypothesis outperforms
the Poisson hypothesis and the same holds for the Binomial hypothesis.
Table 1 Total variation distance and Bhattacharyya distance between the true distribution h−
steps ahead and the predicted one, calculated for N=1000 simulations for data generated from
INAR(1) process with Negative Binomial innovations. The best results are in bold.
Parameters
α = 0.2, r = 1.3
p = 0.3
α = 0.5, r = 1.3
p = 0.3
α = 0.8, r = 1.3
p = 0.3
α = 0.2, r = 12
p = 0.8
α = 0.5, r = 12
p = 0.8
α = 0.8, r = 12
p = 0.8
Parameters
α = 0.2, r = 1.3
p = 0.3
α = 0.5, r = 1.3
p = 0.3
α = 0.8, r = 1.3
p = 0.3
α = 0.2, r = 12
p = 0.8
α = 0.5, r = 12
p = 0.8
α = 0.8, r = 12
p = 0.8
Total variation distance
Model
n=50
n=100
h=1 h=2 h=1 h=2
Negative Bin 0.085 0.080 0.060 0.056
Poisson 0.262 0.261 0.261 0.259
Negative Bin 0.087 0.091 0.060 0.064
Poisson 0.217 0.231 0.209 0.223
Negative Bin 0.086 0.096 0.061 0.067
Poisson 0.171 0.185 0.165 0.177
Negative Bin 0.076 0.071 0.053 0.050
Poisson 0.080 0.077 0.065 0.063
Negative Bin 0.073 0.078 0.053 0.058
Poisson 0.073 0.079 0.057 0.064
Negative Bin 0.074 0.084 0.052 0.060
Poisson 0.070 0.081 0.051 0.059
n=300
h=1 h=2
0.034 0.031
0.257 0.254
0.033 0.036
0.203 0.218
0.034 0.037
0.161 0.170
0.033 0.031
0.053 0.052
0.031 0.034
0.042 0.047
0.031 0.035
0.034 0.039
Bhattacharyya distance
Model
n=50
n=100
h=1 h=2 h=1 h=2
Negative Bin 0.0077 0.0071 0.0037 0.0032
Poisson 0.0660 0.0658 0.0635 0.0630
Negative Bin 0.0089 0.0101 0.0043 0.0050
Poisson 0.0491 0.0545 0.0471 0.0523
Negative Bin 0.0086 0.0111 0.0040 0.0050
Poisson 0.0314 0.0377 0.0300 0.0357
Negative Bin 0.0064 0.0055 0.0037 0.0030
Poisson 0.0073 0.0064 0.0050 0.0043
Negative Bin 0.0073 0.0084 0.0036 0.0042
Poisson 0.0070 0.0083 0.0040 0.0047
Negative Bin 0.0064 0.0085 0.0030 0.0039
Poisson 0.0060 0.0083 0.0031 0.0042
n=300
h=1 h=2
0.0012 0.0011
0.0617 0.0615
0.0013 0.0015
0.0427 0.0480
0.0013 0.0017
0.0276 0.0322
0.0012 0.0010
0.0029 0.0029
0.0012 0.0014
0.0020 0.0024
0.0011 0.0015
0.0015 0.0018
6
Luisa Bisaglia, Margherita Gerolimetto and Paolo Gorgi
Varianza stimatore alfa
●
−0.01
●
mse.a
−0.03
d.a
●
−0.05
●
100
200
300
400
500
600
0.000 0.005 0.010 0.015
Distorsione stimatore alfa
●
●
●
●
700
100
200
300
n
500
600
700
600
700
600
700
Varianza stimatore r
5
400
d.r
mse.r
10
●
●
800
1200
Distorsione stimatore r
15
400
n
●
●
100
200
300
●
400
500
600
●
0
0
●
700
100
200
300
n
mse.p
●
●
0.00
●
300
400
n
500
600
700
0.00 0.01 0.02 0.03 0.04
0.08
d.p
0.04
Varianza stimatore p
●
200
500
n
Distorsione stimatore p
100
●
400
●
●
●
●
100
200
300
400
500
n
Fig. 1 DGP: NBINAR(1). On the left side of the panel the bias of αˆ and m (with the increase of
the sample size) is reported, on the right side the variance (confidence level 95%, 1000 simulations
with α = 0.5, r = 6, p = 0.5).
References
Al-Osh, M. A. & Alzaid, A. A. (1987). First order integer-valued autoregressive
INAR(1) process. Journal of Time Series Analysis, 8(3), 261–275.
Freeland, R. K. & McCabe, B. P. M. (2004). Forecasting discrete valued low count
time series. International Journal of Forecasting, 20, 427–434.
Jung, R. C., G., R., & Tremayne, A. R. (2005). Estimation in conditional first order
autoregression with discrete support. Statistical Papers, 46, 195–224.
McKenzie, E. (1988). Some ARMA models for dependent sequences of poisson
counts. Advances in Applied Probability, 20, 822–835.